Our systematic review was designed and implemented based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (20). The study was registered in Prospero (CRD42022362066).

In the initial screening, two researchers (XYQ and CG) independently searched Embase, Web of Science, Medline, PubMed, and Cochrane Library databases and collected all studies from the date of database creation to March 2023. The main search terms were “osteoarthritis,” “osteoarthrosis,” “osteoarthritides,” and “platelet-rich plasma.” In addition, we manually searched for other relevant literature, such as studies included in some systematic reviews and meta-analyses, to broaden the search for eligible articles. For instance, the following search strategy was used for PubMed: ((“Osteoarthritis”[Mesh]) OR (((((((Osteoarthritides) OR (Osteoarthrosis)) OR (Osteoarthroses)) OR (Degenerative Arthritides[Title/Abstract])) OR (Degenerative Arthritis[Title/Abstract])) OR (Arthrosis[Title/Abstract])) OR (Arthroses[Title/Abstract]))) AND ((“Platelet-Rich Plasma”[Mesh]) OR ((Plasma, Platelet-Rich) OR (Platelet Rich Plasma))). A similar search strategy was used for the other databases. Full details of the search strategy for all databases can be found in Data Sheet 1.

The inclusion and exclusion criteria for literature screening were predetermined to allow for a more rigorous process. Inclusion criteria included (1) patients diagnosed with OA by clinical examination; (2) trial group received PRP injection treatment intervention; (3) control group received HA injection treatment or saline injection treatment; (4) outcome indicators of functional activity and analgesic effects were assessed by relevant scales, such as the Visual Analog Scale (VAS), Western Ontario and McMaster Universities Arthritis Index (WOMAC), Knee Injury and Osteoarthritis Outcome Score (KOOS), and International Knee documentation Committee (IKDC); (5) the experimental design was a RCT; (6) the language was restricted to English.

The exclusion criteria were as follows: (1) animal experiments; (2) IA injection of other drugs, such as Procaine and Lidocaine, over the previous 1 year; (3) full-text content not available; and (4) missing or duplicate experimental data.

We imported all retrieved studies into the document management software Endnote 20 and removed duplicate studies. Two researchers (YX and CG) read the title and abstract of each study simultaneously and screened the studies based on previously developed criteria. For further screening, two reviewers downloaded and read the full texts, removed articles that did not meet the inclusion criteria, and discussed them to confirm their eligibility. If there was a disagreement about the screening process for a particular study, a consensus was reached through advice provided by the principal investigator (WL).

Two reviewers (YX and CG) independently extracted the following data from the included literature: first author of the study, year of publication, sample size, patient age, sex, disease type, disease duration, interventions, and outcome indicators of the trial. Specific therapeutic parameters, including the treatment site, dosage, and number of injections, were also recorded in detail for the intervention method. In addition, when the two reviewers (YX and CG) encountered unclear or complicated extraction of complete literature, the original authors were contacted to obtain complete experimental data. After three consecutive emails, the study was considered missing data if no response was received from the original author.

A quality assessment of the literature was completed independently by two researchers (YX and CG), followed by a discussion to produce consistent results. A risk of bias assessment was conducted using the Cochrane Risk of Bias tool (Review Manager 5.40). Items were assessed in seven areas: blinding of participants and personnel, blinding of outcome assessments, allocation concealment, random sequence generation, incomplete outcome data, selective reporting, and other biases. Risk of bias was graded as high, low, or unclear (21). The risk of bias for each study was documented by mapping the risk of bias assessment.

Heterogeneity between studies was statistically analyzed using RevMan 5.40. The size of heterogeneity was expressed as I²; high heterogeneity was judged when I² ≥ 75%, moderate heterogeneity when 75% > I² ≥ 50%, low heterogeneity when 50% > I² ≥ 25%, and no heterogeneity when I² = 0% (22).

The quality of evidence for the outcome indicators was assessed using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) system, which examines areas such as study limitations, intermittency, inconsistency, and imprecision of results (23). The results were assessed by grading the evidence for the outcome indicators as “high,” “moderate,” “low,” or “very low,” and the strength of the recommendations was divided into two levels: “strong” and “weak” (24).

For statistical and analytical purposes, the extracted study data were entered into RevMan 5.40 software. This meta-analysis was performed using either a fixed-effects or a random-effects model, depending on the size of heterogeneity. When I² ≥ 50%, a random effects model was used; when I² < 50%, a fixed effects model was used (25). We used 95% confidence intervals (CI) and mean differences (MD) to measure the effect size. The effect sizes of different units of the same outcome index were measured using the standardized mean difference.

The initial search resulted in 5,187 studies obtained from five databases. After screening for duplicate studies using the software deletion function, 4,559 studies remained. Two reviewers (YQX and CG) read the titles and abstracts of these studies and screened 4,491 studies that were not relevant to the topic. The remaining 68 studies were read, leaving 44 excluded studies, including two reviews, four case reports, six study protocols, 18 studies with non-compliant interventions, 8 non-RCT studies, and 6 studies with no data. Twenty–four eligible studies were included in the final analysis (26–49) (Figure 1).

Baseline data and experimental intervention parameters were extracted from the 24 included RCTs, summarized in Tables 1, 2, respectively. A total of 1344 patients with OA were eligible across the 24 studies, 712 of whom received PRP injections, and 683 who only received HA or saline as controls (some KOA patients have received treatment in both knees). The intervention characteristics of PRP treatment, such as injection site, injection dose, presence of leukocytes, number of injections, and frequency of injections, are shown in Table 2. In our meta-analysis, 10 studies with a total of 211 patients received one PRP injection, and the remaining 14 studies with a total of 501 patients received two or more PRP injections.

The risk of bias was assessed using the Cochrane Risk of Bias Assessment Tool (RevMan 5.40) for the 24 included RCTs. In 4 of the 24 studies, investigators did not use reasonable blinding of participants and staff, resulting in a high risk of performance bias. In another 7 of the 24 studies, investigators did not use reasonable blinding when assessing outcome indicators, resulting in a high risk of detection bias. Other studies have shown a low or unclear risk in all risk–bias assessments. Overall, the included RCTs had a low risk of bias (Figures 2, 3).

We assessed the level of evidence for the main outcome indicators (VAS, IKDC, KOOS, and WOMAC) in the included studies by using GRADE. The quality of evidence for IKDC and WOMAC was assessed as moderate owing to high heterogeneity (I² > 80%), with serious inconsistencies between studies. The quality of evidence for the VAS and KOOS was rated as high because no serious risk or downgrading factors were identified in any of the projects. Overall, the GRADE recommendation rating was “strong” for the four primary outcome indicators (Table 3).

The VAS was used by investigators in 17 studies (26, 28–30, 32–35, 40–42, 44–49) to assess pain intensity. In these studies, 966 subjects participated in PRP injection therapy trials, divided into experimental and control groups. The I² value was 87%. Therefore, a random-effects model was used. Data analysis of the forest plot showed that the experimental group with PRP injection as an intervention had a significant improvement in VAS OA patients' pain scores compared to the control group (MD = −1.42, CI = 95% [−1.88, −0.96], I² = 87%, P < 0.05; Figure 4). A subgroup analysis showed that the pain efficacy of PRP in patients with OA was synchronized in both the rest and movement states, and pain relief was consistent in both states (VAS-rest, MD = −0.92, CI = 95% [−1.58, −0.25], I² = 89%, P < 0.05; VAS-movement, MD = −0.82, CI = 95% [−1.56, −0.08], test for subgroup differences, P>0.05, I²= 0%; Figure 5). Additional subgroup analyses based on injection sites showed no statistical difference in efficacy between IA combined with intra-osseous (IO) simultaneous PRP injections and IA PRP injections in improving VAS scores in patients with OA (IA+IO PRP injections, MD= −0.74, CI = 95% [−1.29, −0.18], I² = 61%, P < 0.05; IA PRP injections, MD= −1.43, CI = 95% [−2.18, −0.68], I² = 87%, P < 0.05, test for subgroup differences, P > 0.05, I² = 52.7%; Figure 6). A subgroup analysis based on leukocyte concentration showed that leukocyte-rich (LR) PRP injections were not considered effective for patients' pain symptoms. In contrast, leukocyte-poor (LP) PRP injections were effective in relieving pain (VAS, LR-PRP, MD = −0.81, CI = 95% [−1.65, −0.03], I² = 83%, P = 0.06 > 0.05; LP-PRP, MD = −1.62, CI = 95% [−2.36, −0.88], I² = 92%, P < 0.05; test for subgroup differences, P>0.05, I² = 11.6%; Figure 7).

The results of the forest plot data analysis based on PRP treatments of different joint sites indicate that PRP injection therapy was effective in reducing VAS pain scores in patients with ankle osteoarthritis (AOA), knee osteoarthritis (KOA) and temporomandibular joint osteoarthritis (TMJOA) compared to controls (AOA, MD = −1.15, CI = 95% [−1.74, −0.56], I² = 40%, P < 0.05; KOA, MD = −1.03, CI = 95% [−1.16, −0.9], I² = 87%, P < 0.05; TMJOA, MD = −1.35, CI = 95% [−1.74, −0.97], I² = 92%, P < 0.05; Figure 8). However, there was no statistically significant difference between the experimental and control groups in VAS pain scores of hip osteoarthritis (HOA, MD = −0.27, CI = 95% [−0.8, 0.26], I² = 56%, P > 0.05; Figure 8).

The KOOS was used by investigators in five studies (27, 29, 33, 37, 38) to assess treatment effects. In these studies, 427 subjects participated in PRP injection treatment trials, divided into experimental and control groups. The I² value was calculated to be 0%. Therefore, we used a fixed effects model in this study. Forest plots with KOOS as an outcome indicator showed that the experimental group had significantly improved pain scores, symptom scores, activities of daily living scale (ADL) scores, and Quality of Life scores (QOL) in patients with OA compared with the control group, but no significant improvement was observed in the sport scores (KOOS-pain, MD = 2.77, CI = 95% [0, 5.53], I² = 0%, P < 0.05; KOOS-symptoms, MD = 3.73, CI = 95% [0.76, 6.71], I² = 0%, P < 0.05; KOOS-ADL, MD = 3.61, CI = 95% [0.79, 6.43], I² = 0%, P < 0.05; KOOS-QOL, MD = 4.66, CI = 95% [0.98, 8.35], I² = 29%, P < 0.05; KOOS-sport, MD = 0.48, CI = 95% [−3.02, 3.98], I² = 0%, P > 0.05; Figure 9).

The WOMAC was used by investigators in 11 studies (26, 29, 31, 32, 34–36, 39, 41, 42, 49) to assess joint functional activity. In these studies, 574 subjects participated in PRP injection treatment trials, divided into experimental and control groups. The I² value was 95%. Therefore, a random-effects model was used. The forest plots with WOMAC as an outcome indicator showed that WOMAC scores of patients with OA improved in the experimental group but not the control group (MD = −1.64, CI = 95% [−2.65, −0.64], I² = 95%, P < 0.05; Figure 10). In the experimental group, patients showed significant improvements in joint pain, joint stiffness, and functional activity (WOMAC-pain, MD = −1.08, CI = 95% [−1.62, −0.53], I² = 87%, P < 0.05; WOMAC-stiffness, MD = −1.17, CI = 88% [−1.72, −0.63], I² = 87%, P < 0.05; WOMAC- function, MD = −1.12, CI = 95% [−1.65, −0.58], I² = 87%, P < 0.05; Figure 11).

The IKDC was used by investigators in two studies (31, 43) to assess the functional activity of the joints. In these studies, 137 subjects participated in PRP injection treatment trials and were divided into experimental and control groups. The I² value was 93%. Therefore, a random-effects model was used. The results of the forest plot data analysis with IKDC as an outcome indicator showed that the experimental group receiving PRP demonstrated improved OA patient IKDC scores when compared to the control group (MD = 17.4, CI = 95% [3.26, 31.54], I² = 93%, P < 0.05; Figure 12).

Of the 24 studies included, 11 reported adverse reactions after PRP injections in patients with OA; however, most symptoms were mild and transient, such as pain and swelling at the injection site. Of these 11 studies, 2 showed more adverse reactions in the PRP injection group, with patients presenting with syncope, dizziness, headache, nausea, gastritis, sweating, and tachycardia, in addition to pain and stiffness (26, 30). Only one study reported intense knee pain in two patients, and all adverse events reported were resolved without any sequelae (38). Based on the results reported in all studies, a few patients withdrew from the experimental studies because they experienced excessive adverse effects (34).